A power operation and inspection data management method based on data fusion

Abstract

The application discloses a kind of power operation and inspection data management method based on data fusion, more specifically to power operation and inspection management field, including by obtaining and preprocessing multi-node operation and inspection data, and by feature extraction and space-time fusion, form second operation and inspection characteristic vector, inject it into operation and inspection job knowledge graph, through graph calculation real-time deduction, generate structured feature.A kind of power operation and inspection data management method based on data fusion by introducing multidimensional priority evaluation model, significantly reduce the cognitive load and time cost of management personnel;Through scene template driving, learning modal inter attention weight and combining power operation and inspection ontology mask constraint, reduce the risk of resource waste or shortage caused by inaccurate artificial estimation;Through operation and inspection job knowledge graph and the real-time deduction of graph calculation, avoid the problem of resource conflict caused by artificial planning limitation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of operation and inspection management, and more particularly, to a power operation and inspection data management method based on data fusion. BACKGROUND

[0002] With the development of intelligent power systems, efficient management and deep utilization of massive and multi-source power operation and inspection data have become the key to ensuring the safe and stable operation of power grids and improving operational efficiency. Traditional power operation and inspection systems mainly collect device operation parameters periodically and compare them with historical data for analysis to determine the device operation state and potential fault risks. However, the traditional scheme has strict data barriers between various monitoring systems, operation systems, and resource management systems, and the information fusion capability is severely insufficient, making it difficult to fully integrate and deeply mine multi-source heterogeneous data, ultimately limiting the accuracy of operation and inspection decisions and the timeliness of resource scheduling.

[0003] To overcome the above-mentioned defects, the prior art integrates various data sources such as maintenance logs, weather information, and device information, and combines time-space correlation algorithms and machine learning models for fault prediction or health state evaluation, achieving preliminary convergence of multi-source data in the system architecture, improving data accessibility and analysis efficiency in some specific scenarios.

[0004] However, in actual use, there are still some shortcomings, such as lack of deep fusion and collaborative computing mechanism for multi-source data, which leads to inability to dynamically and quantitatively depict the time-space distribution load of each operation, the emergency degree of dynamic changes, and the correlation calculation complexity faced in global resource scheduling based on real-time data when handling high-concurrency and multi-scale data, directly leading to rough decision support granularity, response delay, and serious constraints on the management and scheduling of operation and inspection personnel. SUMMARY

[0005] To overcome the above-mentioned defects of the prior art, the present application provides a power operation and inspection data management method based on data fusion, which solves the problems raised in the background art through the following scheme.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0007] A power operation and inspection data management method based on data fusion, comprising:

[0008] S1: obtaining first operation and inspection data of a plurality of monitoring nodes in a target area, the first operation and inspection data including asset working conditions, historical operation records, and available manpower and material resource information;

[0009] S2: preprocessing the first operation and inspection data to generate first operation and inspection key data containing a unified time-space stamp;

[0010] S3: performing an operation and inspection feature analysis operation on the first operation and inspection key data, the operation and inspection feature analysis operation being configured to extract a first operation and inspection feature vector representing an operation and inspection state in a target region, the operation and inspection state including a spatiotemporal distribution load of operation and inspection tasks, an urgency of each operation and inspection task, and a correlation calculation complexity required for making a resource scheduling decision;

[0011] S4: based on the first operation and inspection feature vector, performing spatiotemporal superposition correlation analysis by a predictive fusion algorithm to extract a second operation and inspection feature vector representing resources required by each operation and inspection task;

[0012] S5: injecting the second operation and inspection feature vector into a pre-constructed operation and inspection task knowledge graph, and performing real-time deduction by graph calculation to generate a structured third operation and inspection feature vector;

[0013] S6: visualizing the third operation and inspection feature vector, and capturing human-computer interaction feedback of a management personnel for reference by the management personnel.

[0014] Preferably, the S1, the first operation and inspection data is obtained through a unified data access gateway.

[0015] The unified data access gateway is configured to receive multi-source heterogeneous data streams from at least two of a meteorological system, a power grid monitoring system, a geographic information system, a production management system, and a user power consumption information acquisition system.

[0016] Preferably, the S3, before performing the operation and inspection feature analysis operation, comprises:

[0017] Based on the first operation and inspection key data, calculating a spatiotemporal density score, a business urgency score, and a correlation calculation complexity score for application to a pre-constructed multi-dimensional priority evaluation model;

[0018] Inputting the spatiotemporal density score, the business urgency score, and the correlation calculation complexity score into a pre-constructed multi-dimensional priority evaluation model to calculate a comprehensive priority score;

[0019] According to the comprehensive priority score, dynamically cutting a data stream corresponding to the first operation and inspection key data into data slices based on time, space, or equipment dimensions.

[0020] Preferably, the S3, according to the comprehensive priority score, injecting data slices into processing queues of different priorities, and when insufficient downstream consumption capacity is monitored, triggering a back pressure mechanism;

[0021] The back pressure mechanism is configured to dynamically reduce a data access rate, and store part of low-priority data slices to a warm data landing disk, and load the data slices again after the pressure is relieved;

[0022] The downstream consumer capacity deficiency judgment condition is that the processing queue length continuously exceeds 1000 elements or the consumer delay exceeds 5 seconds.

[0023] Preferably, the S4, the predictive fusion algorithm is driven by a scene template.

[0024] The scene template predefines a target fusion scene, a required feature mode list, an algorithm combination of feature fusion and an execution order thereof, and dimensions and semantics of a second operation and inspection feature vector.

[0025] Preferably, the S4, extracting the second operation and inspection feature vector specifically includes:

[0026] calculating an attention weight matrix between feature vectors;

[0027] masking and constraining the attention weight matrix by using a predefined power operation and inspection ontology relationship;

[0028] performing weighted summation using the masked attention weight to generate a fusion feature vector representing risk, state and correlation, that is, a second operation and inspection feature vector.

[0029] Preferably, the S4, calculating the attention weight matrix between feature vectors specifically includes:

[0030] learning a query vector corresponding to the feature vector of each mode , a key vector , and a value vector for each mode , wherein is a trainable projection matrix exclusive to each mode , and

[0031] using scaled dot product attention to calculate the similarity between of a mode and of all modes, specifically represented as:

[0032] ,

[0033] wherein is a similarity score of the mode to the mode , , is a scaling factor used to prevent gradient disappearance caused by too large dot product results;

[0034] applying a mask to the similarity score of the mode to the mode obtainable masking converted into a probability distribution, i.e.

[0035] Preferably, the S5, performing the real-time deduction, specifically includes:

[0036] Taking the second operation and inspection feature vector as a trigger point, a relevant subgraph is extracted in the operation and inspection job knowledge graph;

[0037] Based on the predefined semantic rules and the graph neural network model, the subgraph is mixed inferred.

[0038] Technical effects and advantages of the present application:

[0039] 1. The present application realizes the automatic quantitative evaluation of the operation and inspection state by introducing a multi-dimensional priority evaluation model, significantly reduces the cognitive load and time cost of the management personnel, and provides standardized and consistent data basis for the management and scheduling of the operation and inspection personnel;

[0040] 2. The present application drives the scene template, learns the attention weight between modes, and combines the power operation and inspection ontology mask constraint, which excludes false associations that do not conform to the business logic, significantly improves the accuracy and reliability of resource prediction, reduces the risk of resource waste or shortage caused by inaccurate manual estimation, and reduces the uncertainty of the management personnel in the planning stage;

[0041] 3. The present application realizes the automatic solution and optimization of complex resource scheduling problems through real-time deduction of the operation and inspection job knowledge graph and graph calculation, greatly improves the efficiency and rationality of resource scheduling, and avoids the problem of resource conflict caused by the limitations of manual planning. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 A step block diagram of a power operation and inspection data management method based on data fusion according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0044] The terminology used in the following embodiments of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used in the description of the application, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0045] Hereinafter, the terms "first", "second", and "third" are used only for the purpose of description, and cannot be understood as implying or suggesting relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second", and "third" can explicitly or implicitly include one or more of the features, and in the description of the embodiments of the present application, the meaning of "a plurality of" is two or more, unless otherwise specified.

[0046] As shown in the accompanying drawings Figure 1 A power operation and inspection data management method based on data fusion, by acquiring and preprocessing multi-node operation and inspection data, and through feature extraction and spatio-temporal fusion, forming a second operation and inspection feature vector, injecting it into an operation and inspection job knowledge graph, through graph calculation real-time deduction, generating structured features; Specifically, the following steps are included:

[0047] S1: acquiring first operation and inspection data of a plurality of monitoring nodes in a target area, the first operation and inspection data including asset working conditions, historical operation records, and available manpower and material resource information;

[0048] S2: preprocessing the first operation and inspection data to generate first operation and inspection key data containing a unified spatio-temporal stamp;

[0049] S3: performing operation and inspection feature analysis operation on the first operation and inspection key data, the operation and inspection feature analysis operation being used to extract a first operation and inspection feature vector representing the operation and inspection state in the target area, the operation and inspection state including the spatio-temporal distribution load of the operation and inspection job, the urgency of each operation and inspection job, and the associated computational complexity required for resource scheduling decision;

[0050] S4: based on the first operation and inspection feature vector, performing spatio-temporal superposition correlation analysis through a predictive fusion algorithm to extract a second operation and inspection feature vector representing the resources required for each operation and inspection job;

[0051] S5: injecting the second operation and inspection feature vector into a pre-constructed operation and inspection job knowledge graph, and performing real-time deduction through graph calculation to generate a structured third operation and inspection feature vector;

[0052] S6: visualizing the third operation and inspection feature vector and capturing human-computer interaction feedback of the management personnel for reference by the management personnel.

[0053] Specifically, in S1, the first operation and inspection data is acquired through a unified data access gateway; the unified data access gateway is configured with a parsing plug-in corresponding to a data source type, and is used to receive a multi-source heterogeneous data stream from at least two of a meteorological system, a power grid monitoring system, a geographic information system, a production management system and a user power consumption information acquisition system.

[0054] As a specific implementation example, in the scenario of mountainous power transmission line geological disaster risk early warning under heavy rain, data from multiple monitoring nodes, which are of different sources and formats, need to be analyzed collaboratively, and the data is the first operation and inspection data.

[0055] Further, the first operation and inspection data can specifically include but is not limited to: data from the meteorological system: real-time precipitation grid data from a meteorological department, short-term and medium-term wind speed or wind direction prediction data, etc.; data from the power grid monitoring system: line load and switch state data from a SCADA system, equipment state online monitoring data, etc.; the equipment state online monitoring data in this embodiment includes but is not limited to transformer oil chromatography, GIS micro water monitoring, etc.; data from the geographic information system: vector layers containing terrain slope, geological information, power transmission line tower geographic coordinates; data from the production management system: asset working condition, historical operation record, available manpower and material resource information, etc. structured data; data from the user power consumption information acquisition system: load data, voltage curve, etc.

[0056] It should be noted that the unified data access gateway supports multiple communication protocol interfaces, including but not limited to HTTP / HTTPS, MQTT, Kafka, FTP / SFTP, JDBC, to adapt to the push and pull modes of different data sources; in the embodiment, the data of the meteorological system is pushed to the gateway through the MQTT protocol, while the historical account data of the production management system is pulled by the gateway at regular intervals through the JDBC connection; the parsing plug-in complies with a predefined interface specification, which requires the parsing plug-in to implement a self-identification method and a parsing method, and can be dynamically loaded and unloaded by the plug-in management service in the gateway; the self-identification method determines whether to process the data stream by matching specific features of the data stream, including but not limited to: specific identifiers contained in the data packet header, agreed transmission protocol ports, path patterns of data source URLs, or magic numbers in the first few bytes of raw data; in the embodiment, the plug-in for parsing the JSON data of the provincial meteorological bureau is configured to check whether the URL of the HTTP request contains a preset path; the workflow of the parsing plug-in includes: calling the self-identification method to determine whether the input data stream can be processed, and if matched, calling the parsing method to convert the raw unstructured or semi-structured data into intermediate structured data, which at least includes: a unique identifier of the data source, a copy of the raw data, valid information extracted from the raw data, a timestamp of when the gateway received the data, an event occurrence timestamp parsed from the raw data, raw spatial coordinate information and coordinate system identifier parsed from the raw data, and other information.

[0057] Further, the unified data access gateway also includes a traffic control and exception handling function; when detecting a data source connection interruption, data format exception or traffic overload, an error log is recorded and a retry, degradation or alarm is performed according to a preconfigured strategy; at the same time, a runtime state monitoring interface is provided to output the connection state, data receiving rate and parsing success rate indicators of each data source in real time.

[0058] Specifically, in S2, the first operation and inspection data is converted into directly usable high-quality, standardized data, i.e., first operation and inspection key data; the preprocessing includes a time-space reference alignment operation and a semantic standardization operation, and the implementation steps are as follows: the time-space reference alignment operation, the time information carried in the intermediate structured data is uniformly converted into a world coordinated time timestamp, and the geographic space information is uniformly converted into a WGS84 coordinate system; the semantic standardization operation, by a semantic rule engine, at least one of unit conversion, enumeration value mapping or numerical normalization is performed on the intermediate structured data to generate the first operation and inspection key data; the first operation and inspection key data is output in a standardized multi-dimensional data object format; the multi-dimensional data object at least includes a UTC timestamp, a latitude and longitude structure in a WGS84 coordinate system, a data type field and an attribute field in the form of a key-value pair.

[0059] It should be noted that the time-space reference alignment operation is performed by a dedicated time-space processing submodule; the time-space processing submodule has a time zone database and a coordinate conversion library built-in; for time alignment, according to the time zone information carried in the first operation and inspection data, various local times are converted into accurate UTC timestamps; for space alignment, the coordinate information in the first operation and inspection data is identified, and a coordinate conversion function is called to convert it into latitude and longitude coordinates in the WGS84 coordinate system; the workflow of the semantic rule engine is as follows: rule matching: receiving the time-space aligned data, and matching the preloaded corresponding semantic rule set according to the data source type and data structure; rule execution: executing the matched rules in priority order, the rules are composed of conditions and actions; the condition is used to judge whether the data field meets a specific pattern, and the action is used to perform field mapping, unit conversion, enumeration value mapping or numerical normalization operations; result output: after all rules are executed, the final converted data is output.

[0060] Specifically, in S3, the first operation and inspection key data output by S2 is received, and a resource scheduler is introduced to perform operation and inspection feature analysis operation.

[0061] In a possible implementation, before performing the operation and inspection feature analysis operation, the method further includes: based on the first operation and inspection key data, calculating a time-space density score, a service emergency degree score, and an associated calculation complexity score of a time-space distribution load of operation and inspection work, an emergency degree of each operation and inspection work, and an operation and inspection state required for making a resource scheduling decision, to be applied to a pre-constructed multi-dimensional priority evaluation model; inputting the time-space density score, the service emergency degree score, and the associated calculation complexity score into a pre-constructed multi-dimensional priority evaluation model to calculate a comprehensive priority score; and according to the comprehensive priority score, dynamically cutting a data stream corresponding to the first operation and inspection key data into data slices based on a time dimension, a space dimension, or a device dimension; and assigning a corresponding calculation resource and storage strategy to each data slice and scheduling the data slice to a processing queue of different priority, where the storage strategy includes hot data caching, warm data archiving, and cold data archiving.

[0062] It should be noted that the operation of dynamically cutting a plurality of data slices is performed based on one or more of the following dimensions: a time dimension, a space dimension, and a device dimension. The time dimension is cutting according to a fixed time window; the space dimension is cutting according to a geographical administrative region, a power grid dispatching region, or a geographical fence defined by a person; and the device dimension is cutting according to a voltage level, a device type, or a line ownership. The assigned calculation resource refers to scheduling a calculation task to a GPU node group and applying corresponding CUDA core and memory resources.

[0063] In this embodiment, the multi-dimensional priority evaluation model calculates a comprehensive priority score , which is specifically represented as:

[0064] ,

[0065] wherein, is a time-space density score, is a service emergency degree score, is an associated calculation complexity score, are configurable weight coefficients of the time-space density score , the service emergency degree score , and the associated calculation complexity score , and satisfy

[0066] It should be noted that the time-space density score is calculated based on time-space attributes, and is obtained based on a time update frequency and a spatial coverage range of the first operation and inspection key data, and is specifically represented as: is specifically represented as: The service emergency degree score The correlation calculation complexity score is obtained by querying a predefined business rule library for query matching For estimating the amount of computing resources required for processing data, the size of the data unit is multiplied by the complexity level of the algorithm required for processing the data to obtain; for Data slices with a priority score greater than 0.7 are preferentially processed, GPU computing resources and high-priority threads are allocated, more computing resources are allocated, and data slices are accessed in milliseconds through hot data caching; for data slices with a priority score of 0.3 ≤0.7, normal processing is performed, standard CPU computing resources and normal threads are allocated, and data slices are stored on a warm data disk. Data slices with a priority score less than 0.3 are batch processed and stored in cold data archives.

[0067] Further, the resource scheduler injects data slices into different priority processing queues according to their When it is monitored that the downstream consumption capacity is insufficient, a back pressure mechanism is triggered; the judgment condition for the insufficient downstream consumption capacity is that the length of the processing queue continuously exceeds 1000 elements or the consumer delay exceeds 5 seconds; the back pressure mechanism is configured to dynamically reduce the data access rate and store part of the low-priority data slices to the warm data disk, and then load them when the system pressure is relieved.

[0068] Further, the data slices scheduled to the computing resources are executed to generate the first operation and inspection feature vector; in this embodiment, the first operation and inspection feature vector is a floating point number array, the dimension and semantics of which are determined by the algorithm applied by the operation and inspection feature analysis operation, and the first operation and inspection feature vector is packaged with the unique identifier of the corresponding data slice, the comprehensive priority score and the timestamp.

[0069] Specifically, in S4, the predictive fusion algorithm is driven by a scene template; the scene template predefines the target fusion scene, the required feature modal list, the algorithm combination and execution order of feature fusion, and the dimension and semantics of the second operation and inspection feature vector; according to the predefinition of the scene template, the required modal data is actively pulled from the data bus, and the special encoder matched with each modal is called for further feature representation.

[0070] In this embodiment, a convolutional neural network encoder is used to process meteorological grid data to capture spatial distribution features, a long short-term memory network encoder is used to process device time series data to capture time dependence, and a graph neural network encoder is used to process geographic vector data to learn topological relationship features.

[0071] In one possible implementation, the extraction of the second operation and maintenance feature vector is accomplished by a cross-modal attention fusion module, including: receiving feature vectors output from dedicated encoders that match each modality; calculating an attention weight matrix between feature vectors corresponding to different modalities using a learnable weight matrix; applying a mask constraint to the attention weight matrix using predefined power operation and maintenance ontology relationships; and performing a weighted summation using the masked attention weights to generate a fusion feature vector representing risk, state, and correlation, i.e., the second operation and maintenance feature vector.

[0072] It should be noted that this is achieved through a learnable weight matrix, whereby the feature vector of each modality learns its corresponding query vector. Key vector Value vector It means that, among them, Represented as modality, This represents a trainable projection matrix exclusive to each modality. Represented as the feature vector output by a dedicated encoder that matches each modality; a modality's feature vector is computed using scaled dot product attention. With all modalities The similarity between them is specifically expressed as:

[0073] ,

[0074] in, Represented as modality For modes Similarity score, , Represented as a scaling factor to prevent the gradient from vanishing due to excessively large dot product results; a predefined mask matrix based on the power operation and maintenance ontology. Based on the generation of a pre-built knowledge graph ontology of operation and maintenance tasks, its elements ,in Represented as modality With mode Relationships between entities are allowed to exist within the ontology. Represented as modality With mode Relationships between entities are prohibited by the ontology, further applying masks to modalities. For modes similarity score Up, you can get , after the mask Transform into a probability distribution, i.e., attention weights; use the obtained attention weights for all modalities. We perform a weighted summation to obtain a new representation of each modality after interaction, which is the second operational feature vector.

[0075] Further, after generating the second operation and inspection feature vector, a domain rule engine is used for verification; the domain rule engine has a set of Boolean expression rules built-in for checking the rationality of the feature vector; if the feature vector violates the rules, a correction gradient signal is generated to fine-tune the parameters of the cross-modal attention fusion module through back propagation until the output result meets all domain rule constraints; in this embodiment, the Boolean expression rules are used to express domain knowledge constraints; when the Boolean expression rule is IF (meteorology. wind speed > 30 m / s) THEN (device. wind deflection risk index > 0.7), if the output feature vector does not meet this rule, it is determined that the rule is violated.

[0076] Still further, the second operation and inspection feature vector is accompanied by an explainability document; the explainability document records the input feature modality with the highest contribution and its specific attention weight allocation in a structured form during the fusion process.

[0077] Specifically, in S5, the pre-constructed operation and inspection job knowledge graph takes the device, environmental phenomenon, fault, and operation and maintenance measure as the core entity ontology framework, and defines the object attributes of the causal, spatial, and attribute correlations between the core entities.

[0078] It should be noted that the operation and inspection job knowledge graph is constructed by extracting entities and relationships from multi-source unstructured texts such as asset working conditions, historical operation records, available manpower and material resource information, etc. using natural language processing technology; the operation and inspection job knowledge graph supports dynamic updating; the deduction results of S5 and the feedback signals captured in S6 will be used as new knowledge sources to add or delete entities, relationships or adjust relationship weights in the operation and inspection job knowledge graph, realizing the self-evolution of the operation and inspection job knowledge graph

[0079] In a possible implementation, performing the real-time deduction includes: taking the second operation and inspection feature vector as a trigger point to extract a related subgraph in the operation and inspection job knowledge graph; performing hybrid reasoning on the subgraph based on a pre-defined semantic rule and a graph neural network model; the result generated by the hybrid reasoning is a potential event chain; the potential event chain represents an inference path from the current state to a potential consequence; and the entities and relationships in the potential event chain are each accompanied by a probability value, which represents the confidence of each inference relationship in the event chain, output by the Softmax function of the last layer of the graph neural network model.

[0080] It should be noted that the process of extracting the related subgraph is performed by a graph query engine: taking the core entity identified in the second operation and inspection feature vector as a starting point, performing multi-hop traversal in the operation and inspection job knowledge graph to extract all associated entities and relationships to form the related subgraph;

[0081] Further, the third operation and inspection feature vector is a structured decision object; the decision object at least includes a deduced root cause, a potential event chain, a list of recommended operation and maintenance measures, and an overall confidence.

[0082] Specifically, in S6, the visualization is presented on a GIS map base, and the decision information contained in the third operation and inspection feature vector is presented in a layered rendering manner; the decision information includes at least one of device risk level, early warning event chain, and recommended inspection path; the human-computer interaction feedback includes three levels of operation granularity: confirmation or ignore operation on early warning events, drag and adjust operation on recommended operation and maintenance scheme, and operation of injecting new knowledge through forms or annotations; the captured interaction feedback signal is converted into structured training data or knowledge graph incremental information, and is used as a new data source backflow to the data access end of S2; the backflow feedback data will be used to optimize the parameters in the predictive fusion algorithm in S4 and / or the operation and inspection job knowledge graph in S5.

[0083] In the embodiment, when multiple high-risk devices are spatially aggregated, icon aggregation rendering is automatically enabled; when a user selects a single early warning event, related entities are automatically highlighted and irrelevant entities are faded.

[0084] Secondly, in the drawings of the disclosed embodiments, only structures related to the disclosed embodiments are involved, other structures can be referred to the general design, and in the case of no conflict, the same embodiments and different embodiments of the present application can be combined with each other;

[0085] Finally, the above only describes the preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A data fusion-based power operation and inspection data management method, characterized in that, The method comprises the following steps: S1: obtaining first operation and inspection data of a plurality of monitoring nodes in a target area, wherein the first operation and inspection data comprises asset working conditions, historical operation records, and available manpower and material resource information; S2: preprocessing the first operation and inspection data to generate first operation and inspection key data containing a unified space-time stamp; S3: performing operation and inspection feature analysis on the first operation and inspection key data, wherein the operation and inspection feature analysis is used to extract a first operation and inspection feature vector representing the operation and inspection state in the target area, and the operation and inspection state comprises the space-time distribution load of operation and inspection work, the emergency degree of each operation and inspection work, and the associated calculation complexity required for resource scheduling decision-making; Before performing the operation and inspection feature analysis, the method comprises the following steps: based on the first operation and inspection key data, calculating the space-time density score, the business emergency degree score, and the associated calculation complexity score for application to a pre-constructed multi-dimensional priority evaluation model; inputting the space-time density score, the business emergency degree score, and the associated calculation complexity score into a pre-constructed multi-dimensional priority evaluation model to calculate a comprehensive priority score; and according to the comprehensive priority score, dynamically cutting the data stream corresponding to the first operation and inspection key data into data slices based on the time, space, or equipment dimensions; S4: based on the first operation and inspection feature vector, performing space-time superposition correlation analysis by a predictive fusion algorithm to extract a second operation and inspection feature vector representing the required resources of each operation and inspection work; The predictive fusion algorithm is driven by a scene template; the scene template predefines a target fusion scene, a required feature mode list, an algorithm combination and an execution order of feature fusion, and a dimension and semantics of the second operation and inspection feature vector; according to the predefinition of the scene template, the required modal data is actively pulled from the data bus, and a special encoder matched with each modal is called for further feature representation; S5: injecting the second operation and inspection feature vector into a pre-constructed operation and inspection work knowledge graph, performing real-time deduction by graph calculation to generate a structured third operation and inspection feature vector; Performing the real-time deduction specifically comprises: taking the second operation and inspection feature vector as a trigger point to extract a related subgraph in the operation and inspection work knowledge graph; and performing hybrid reasoning on the subgraph based on a pre-defined semantic rule and a graph neural network model; S6: visualizing the third operation and inspection feature vector and capturing human-computer interaction feedback of a management personnel.

2. The power operation inspection data management method based on data fusion according to claim 1, characterized in that: The first operation and inspection data is obtained through a unified data access gateway; The unified data access gateway is used to receive multi-source heterogeneous data streams from at least two of a meteorological system, a power grid monitoring system, a geographic information system, a production management system, and a user power consumption information acquisition system.

3. The power operation inspection data management method based on data fusion according to claim 1, characterized in that: According to the comprehensive priority score, the data slices are injected into different priority processing queues, and when it is monitored that the downstream consumption capacity is insufficient, a back pressure mechanism is triggered; The back pressure mechanism is configured to dynamically reduce the data access rate and store part of the low-priority data slices to a warm data landing disk, and then load them after the pressure is relieved. The downstream consumption capability insufficient judgment condition is that the processing queue length continuously exceeds 1000 elements or the consumer delay exceeds 5 seconds.

4. The power operation inspection data management method based on data fusion according to claim 1, characterized in that: The S4, extracting the second operation and inspection feature vector, specifically comprises: By calculating the attention weight matrix between feature vectors; Using a predefined power operation and inspection ontology relationship to mask constrain the attention weight matrix; Using the masked attention weight for weighted summation to generate a fusion feature vector representing risk, state and correlation, namely the second operation and inspection feature vector.

5. The power operation inspection data management method based on data fusion according to claim 4, characterized in that: The S4, calculating the attention weight matrix between feature vectors, specifically comprises: By calculating the attention weight matrix between feature vectors; learning to each modality's feature vector its corresponding query vector key vector value vector denotes, wherein, denotes for a modality, denotes a trainable projection matrix exclusive to each modality, denotes a feature vector output by a dedicated encoder that matches each modality The similarity between the scaled dot-product attention computed for one modality and all modalities is computed as follows, specifically: ， wherein, is represented as a modality a similarity score for a modality , , is represented as a scaling factor to prevent the dot product result from being too large to cause gradient vanishing; Applying masks to modalities Similarity scores for modalities Obtain Convert masked to probability distributions, i.e. attention weights.​

Images